Research map / 3D vision / Depth and stereo
CroCo: Self-Supervised Pre-training for 3D Vision Tasks by Cross-View Completion
Philippe Weinzaepfel, Vincent Leroy, Thomas Lucas, Romain Brégier, Yohann Cabon, Vaibhav ARORA, Leonid Antsfeld, Boris Chidlovskii, Gabriela Csurka, Jerome Revaud
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| cross-view completion pre-training instead of single-view MAE on the same Habitat data (ablation) | ViT-Base/16 · NYUv2 depth | delta_1: 79.0 -> 85.6 | Improved |
| cross-view completion instead of MAE on same Habitat data (ablation) | ViT-Base/16 · Taskonomy average L1 | L1 x1000: 35.65 -> 33.00 | Improved |
| cross-view completion instead of MAE on same Habitat data (ablation) | ViT-Base/16 · ADE20k segmentation | mIoU: 40.3 -> 40.6 | No clear effect |
| normalize reconstruction targets per patch (ablation) | CroCo CrossBlock · NYUv2, ADE, Taskonomy | delta_1 / mIoU / L1: 83.4 -> 85.6 (NYUv2); 39.0 -> 40.6 (ADE); 34.74 -> 33.00 (Taskonomy) | Improved |
| use CatBlock instead of CrossBlock decoder (ablation) | CroCo · NYUv2 / ADE / Taskonomy | delta_1 / mIoU / L1: 85.6 -> 86.2 (NYUv2); 33.00 -> 33.35 (Taskonomy) | Mixed |
| use two real viewpoints instead of geometric transformations of one image to build pairs (ablation) | CroCo · NYUv2 depth | delta_1: 66.1 -> 86.8 | Improved |
| use 90% masking ratio (ablation) | CroCo CrossBlock · ADE, NYUv2, Taskonomy | downstream performance | Improved |
| keep pre-trained decoder when finetuning on monocular tasks (ablation) | CroCo · NYUv2 depth | delta_1: 86.1 -> 88.1 | Improved |
| pre-train on Habitat indoor scenes instead of ImageNet-1K | MAE · ImageNet-1K linear probing | top-1 accuracy: 68.0 -> 32.5 | Hurt |
| initialize with CroCo encoder and decoder instead of MAE (Habitat) encoder | optical flow regression · MPI-Sintel clean | AEPE: 4.63 -> 3.00 | Improved |
| CroCo pre-training instead of random init | optical flow regression · MPI-Sintel clean | AEPE: 18.81 -> 3.00 | Improved |
| CroCo init instead of MAE (Habitat) encoder with scratch decoder | relative pose regression · 7-scenes | median position error: 24.8cm -> 5.0cm | Improved |
Closest papers by meaning
- CroCo v2: Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow (ICCV 2023 · similarity 0.93)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.88)
- Cross-view and Cross-pose Completion for 3D Human Understanding (CVPR 2024 · similarity 0.88)
- Masked Scene Modeling: Narrowing the Gap Between Supervised and Self-Supervised Learning in 3D Scene Understanding (CVPR 2025 · similarity 0.87)
- MonoMAE: Enhancing Monocular 3D Detection through Depth-Aware Masked Autoencoders (NeurIPS 2024 · similarity 0.87)
- Cross-View Completion Models are Zero-shot Correspondence Estimators (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
